HIV Protease Drug Discovery

Scaffold-aware HIV-1 protease potency modeling and compound prioritization

139 curated structures · scaffold-aware validation · uncertainty and domain checks · one-command reproducibility

---

Table of contents

• Overview
• Executive snapshot
• Dataset
• Workflow
• Evidence
• Project structure
• Quick start
• Limitations
• Author and license

---

Overview

HIV Protease Drug Discovery turns public HIV-1 protease bioactivity records into a transparent compound-review queue. It standardizes molecular structures, prevents scaffold leakage, benchmarks a fingerprint model, surfaces prediction uncertainty, checks chemical-domain coverage and simulates a retrospective design–test–learn loop.

The emphasis is not a headline model score. It is the quality of the decision evidence: what should be reviewed next, how confident is the model, and where can it fail?

Executive snapshot

• Curated 139 unique HIV-1 protease structures from exact ChEMBL IC50 records.
• Evaluated unseen chemistry using a 107/32 scaffold split, not a random row split.
• Reduced holdout RMSE from 0.990 (mean baseline) to 0.700 pActivity units.
• Delivered an auditable ranking with potency, uncertainty, similarity and molecular-property evidence.
• Kept docking and experimental claims explicitly outside the validated scope.

Dataset

The bundled sample is derived from public ChEMBL activity records for HIV-1 protease (CHEMBL243). Only exact, standardized IC50 values reported in nanomolar units are retained; repeated canonical structures are aggregated by median pActivity.

Quality control	Result
Curated structures	139
Duplicate canonical SMILES	0
Missing SMILES / pActivity	0 / 0
Target	HIV-1 protease
Endpoint	IC50 converted to pActivity

See the data dictionary and data provenance.

Workflow

flowchart LR
    A["ChEMBL IC50 records"] --> B["Structure and assay curation"]
    B --> C["Bemis–Murcko scaffold split"]
    C --> D["Morgan fingerprint baselines"]
    D --> E["Uncertainty and domain checks"]
    E --> F["Multi-objective review queue"]
    F --> G["Retrospective label reveal"]
    G --> D

Loading

The model uses 2,048-bit Morgan fingerprints and Extra Trees regression. A mean predictor provides the minimum comparison; tree-level dispersion indicates uncertainty, while nearest-training-set Tanimoto similarity exposes applicability-domain risk. Full choices are documented in the methodology.

Evidence

1. The model beats a naïve baseline on unseen scaffolds

The scaffold holdout achieved R² 0.452, RMSE 0.700 and MAE 0.570, compared with baseline RMSE 0.990. This is useful diagnostic evidence on a bounded sample—not proof of prospective discovery performance.

2. Ranking is decision-aware, not potency-only

Every held-out molecule is ranked without exposing its measured activity. The score combines predicted potency (55%), lower uncertainty (20%), training-domain similarity (15%) and a broad molecular-weight screen (10%). Low-domain predictions are automatically penalized.

Evidence delivered	Decision value
Predicted pActivity	Expected potency signal
Tree dispersion	Relative model uncertainty
Nearest-train similarity	Applicability-domain context
Molecular descriptors	Early chemistry review prompts
Transparent weighted score	Reproducible prioritization

3. The learning loop tests whether selections improve the model

In the retrospective simulation, external-scaffold RMSE moved from 0.873 to 0.736 across three label-reveal rounds. Because the labels already existed in ChEMBL, this demonstrates workflow behavior only—not a lab-in-the-loop result.

Project structure

HIV-Protease-Drug-Discovery/
├── src/hiv_protease_discovery/ # Curation, chemistry, modeling and ranking
├── notebooks/               # Executed analytical narrative
├── data/                    # Bounded public sample and source response
├── results/                 # Regenerable tables and figures
├── docs/                    # Methodology, model card, setup and docking handoff
├── tests/                   # Unit tests for core scientific logic
└── .github/workflows/       # Automated lint and test checks

Quick start

git clone https://github.com/dineshbarri/HIV-Protease-Drug-Discovery.git
cd HIV-Protease-Drug-Discovery
python -m venv .venv

# Windows PowerShell
.venv\Scripts\Activate.ps1

python -m pip install --upgrade pip
pip install -e ".[dev,notebook]"
hiv-protease-discovery demo
pytest

For macOS/Linux activation, notebook launch, data refresh and troubleshooting, use the detailed setup guide.

Limitations

• The dataset is a small, target-specific portfolio sample with heterogeneous public-assay context.
• Scaffold-holdout metrics are more demanding than random splits but remain retrospective estimates.
• Tree dispersion is an uncertainty indicator, not calibrated predictive probability.
• Activity predictions do not establish binding mode, selectivity, ADME, toxicity, efficacy or clinical value.
• The docking-ready handoff is a protocol—not a reported docking experiment.

---

👨‍💻 Author

Dinesh Barri

Data analyst focused on healthcare analytics, business intelligence, Python, SQL and decision-ready storytelling.

Contributions, questions and suggestions are welcome through the repository's Issues tab.

---

📄 License

Source code is available under the MIT License. ChEMBL-derived data remain subject to the original data terms and attribution.

---

Built with analytical care by Dinesh Barri.
_{If you found this project valuable, consider giving it a ⭐.}

Built to make molecular-model decisions reproducible, inspectable and honest.